Cheem viewer
We have extracted tree SHAP, an feature importance measure in the vicinity of each observation. We need to identify an instance of interest to explore; we do so with the linked brushing available in the global view. Then we will vary contributions from different features to test the support an explanation in a radial tour
Global view
To get more complete view lets look at approximations of the data space, attribution space, and model fits side-by-side with linked brushing with the help of plotly and crosstalk. We have identified an observation with a large Mahalanobis distance (in data space) and the closest neighbor in attribution space.
prim <- 1
comp <- 17
global_view(ames_chm, primary_obs = prim, comparison_obs = comp,
height_px = 240, width_px = 720,
as_ggplot = TRUE, color = "log_maha.data")From this global view we want to identify a primary instance (PI) and optionally a comparison instance (CI) to explore. Misclassified or observations with high residuals are good targets for further exploration. One point sticks out in this case. Instance 243 (shown as *) is a Gentoo (purple) penguin, while the model predict it to be a Chinstrap penguin. Penguin 169 (shown as x) is reasonably close by and correctly predicted as Gentoo. In practice we used linked brushing and misclassification information to guide our search.
Radial tour
There is a lot to unpack here. The normalized distribution of all
feature attribution from all instances are shown as parallel coordinates
lines. The above selected PI and CI are shown here as a dashed and
dotted line respectively. The first thing we notice is that the
attribution of the PI is close to it’s (incorrect) prediction of
Chinstrap (orange) in terms of bill length (bl) and flipper
length (fl). In terms of bill depth and body mass
(bd and bm) it is more like its observed
species Gentoo (purple). We select flipper length as the feature to
manipulate.
## Normalized attribution basis of the PI
bas <- sug_basis(ames_xgb_shap, rownum = prim)
## Default feature to manipulate:
#### the feature with largest separation between PI and CI attribution
mv <- sug_manip_var(
ames_xgb_shap, primary_obs = prim, comparison_obs = comp)
## Make the radial tour
ggt <- radial_cheem_tour(
ames_chm, basis = bas, manip_var = mv,
primary_obs = prim, comparison_obs = comp, angle = .15)
## Animate it
animate_gganimate(ggt, fps = 6)
#height = 2, width = 4.5, units = "in", res = 150
## Or as a plotly html widget
#animate_plotly(ggt, fps = 6)
Starting from the attribution projection, this instance already looks more like its observed Gentoo than predicted Chinstrap. However, by frame 8, the basis has a full contribution of flipper length and does look more like the predicted Chinstrap. Looking at the parallel coordinate lines on the basis visual we can see that flipper length has a large gap between PI and CI, lets check the original variables to digest.
library(ggplot2)
prim <- 1
ggplot(penguins_na.rm, aes(x = bill_length_mm,
y = flipper_length_mm,
colour = species,
shape = species)) +
geom_point() +
## Highlight PI, *
geom_point(data = penguins_na.rm[prim, ],
shape = 8, size = 5, alpha = 0.8) +
## Theme, scaling, color, and labels
theme_bw() +
theme(aspect.ratio = 1) +
scale_color_brewer(palette = "Dark2") +
labs(y = "Flipper length [mm]", x = "Bill length [mm]",
color = "Observed species", shape = "Observed species")This profile, with two features that are most distinguished between the PI and CI. This instance is nested in the in between the Chinstrap penguins. That makes this instance particularly hard for a random forest model to classify as decision tree can only make partition on one value (horizontal and vertical lines here).
Shiny application
We provide an interactive shiny application.
Interactive features are made possible with plotly,
crosstalk, and DT. We have
preprocessed simulated and modern datasets for you to explore this
analysis with. Alternatively, bring your own data by saving the return
of cheem_ls() as an rds file. Follow along with the example
in ?cheem_ls.